How fast is ChatGPT’s Ultrafast mode, and is it worth it?

Ultrafast is a speed setting for GPT-6 Astra in Codex and ChatGPT Work. OpenAI says it writes up to 8 times faster than Standard, and it uses your plan’s allowance 8 times as fast. We tested it on 30 September 2026, racing it against Standard on the same coding job and timing single replies. It wrote 4.5 to 14.8 times faster than Standard, depending on what it was writing. A whole job can speed up much less, because a job is more than writing.

We pay for the $500 Pro plan ourselves, and nobody sponsored this guide or the video.

Is Ultrafast worth it for this job?

Put in how long a job takes on Standard and roughly how much of that time the model spends writing. The calculator works out the time on Ultrafast or Fast, and how much more of your allowance it uses.

What Ultrafast is

OpenAI calls Ultrafast a service tier, like Standard and Fast. In Codex and ChatGPT Work it runs GPT-6 Astra, the model OpenAI released on 3 September 2026 as its most capable.

The speed claim is about writing. OpenAI’s Codex docs say “GPT-6 Astra Ultrafast generates tokens up to 8x faster than GPT-6 Astra in Standard mode”, and add: “This comparison measures token generation speed, not billing rates or overall task completion time.” A token is a word or a piece of one. OpenAI’s rule of thumb is that a token is about three quarters of an English word.

Codex, OpenAI’s coding agent, shares its usage with ChatGPT Work, so Ultrafast draws on one allowance in both. OpenAI added Ultrafast for GPT-6 Astra to its API on 29 September 2026.

Who can use it

In ChatGPT, Ultrafast comes with the $500 a month Pro plan and eligible Enterprise and Edu plans. Pro also comes at $100 and $200 a month, and those don’t include it. Other self-serve plans can’t buy their way in either: OpenAI says they “don’t have access to Ultrafast at launch, even with purchased credits.”

On Enterprise it’s off until a workspace owner turns it on, for chosen people or the whole workspace. Older Enterprise plans that run on rate limits instead of usage-based billing can’t use it, and neither can workspaces that need their AI processing kept in a region outside the United States.

Anyone with an OpenAI API key can use it at API prices. OpenAI says it’s “currently available to all API users at low rate limits”. Codex can run on an API key too, and then you pay API token prices instead of using a plan’s allowance.

What it costs

On a plan, you pay for Ultrafast with usage. On Pro $500 it uses your included usage first, then your credits once that runs out. These are OpenAI’s figures, against 1x for Standard:

	Fast	Ultrafast
Speed	about 2x	up to 8x
Included usage	2.5x	8x
Bought credits	2x	6x

Enterprise pay-as-you-go usage is charged at the credit rates. OpenAI points out that “These billing multipliers don’t describe speed increases.”

In the API, at short-context prices, GPT-6 Astra’s output costs $300 per million tokens on Ultrafast, against $50 on Standard and $100 on Fast. Input costs $60, $10 and $20. That’s 6 times Standard, the same as the credit rate.

Fast is the middle option. It’s on the paid plans (Plus, Pro, Business, Enterprise and Edu) and the API, while Free and Go get GPT-6 Luna at Standard speed. Codex’s own model list describes Fast on GPT-6 Astra as “2x speed, increased usage”, and OpenAI’s GPT-6 Astra launch post says up to 2.5 times in the API.

How to turn it on

We used the Codex command line. To make Ultrafast your default there, add these lines to Codex’s config.toml file:

• model = "gpt-6-astra"
• service_tier = "ultrafast"

For a single run, start Codex with codex -m gpt-6-astra -c service_tier="ultrafast" instead.

Codex only offers the speed tiers that OpenAI’s model list gives your account, and ours listed Ultrafast for GPT-6 Astra. OpenAI’s docs list a /fast command that switches Fast mode on and off, but no command for Ultrafast. In the desktop app, the docs say you choose a speed under Advanced. We didn’t test that.

In the API, set the model to gpt-6-astra and service_tier to ultrafast. OpenAI recommends a WebSocket connection, because “Without a persistent connection, network overhead can reduce the latency gains.”

The race

We gave Codex the same job twice, on GPT-6 Astra at medium effort: build a small browser game from the same brief and pictures. One run was on Standard and the other on Ultrafast, and both started at the same moment.

Ultrafast’s first version was done at 57 seconds. We then gave it six more jobs from a list: sound, a cat that knocks mugs off a shelf, a rate rise power-up, a night shift mode, a ghost of your best run and a postcard you can save. It finished them at 1:15, 1:35, 1:50, 2:06, 2:31 and 2:58, and then had nothing left to do. Standard’s first version landed at 8 minutes 31 seconds. Both first versions played start to finish with no errors in our capture. Each run had found and fixed a problem of its own along the way.

For scale, a full 1.7 litre kettle of water at 15°C takes about 3 minutes 57 seconds to boil on a 3 kW British kettle, and about 7 minutes 54 seconds on a 1.5 kW American one such as Breville’s IQ Kettle, if 85% of the power goes into the water. Ultrafast had finished all seven jobs before the British kettle boiled. Standard finished less than a minute after the American one.

The two runs didn’t do identical work. Standard’s first version took 11 model calls and 15,526 output tokens, and Ultrafast’s took 8 calls and 11,331, so the fair comparison is speed within a call. The biggest call in each run wrote the main game file:

Main file	Standard	Ultrafast
Tokens	10,919	7,838
Time	328.8 s	25.8 s
Tokens a second	33.2	304

That’s 9.2 times faster, a little over OpenAI’s “up to 8x”. It’s one day’s measurement, and Standard’s speed can vary with load.

Play the game Ultrafast built

GPT-6 Astra Ultrafast wrote Kettle Dash during the race. Its first version took 57 seconds, and it had added all six upgrades by 2:58. This is that last version, and we haven’t changed a line of its code.

Get Hana to the kettle before it boils. Jump with Space, the up arrow or a tap, and hold to jump higher. Duck with the down arrow or a swipe down. Every bump adds 8°C to the kettle, and the teabags you collect go into a tea jar that earns interest as you run. Sound starts after your first key press or tap, and M turns it off.

The two first versions from the blind test are here too, before any upgrades: Standard’s, finished at 8:31 and Ultrafast’s, finished at 57 seconds.

Speed depends on what it writes

For cleaner figures we timed single replies, three rounds of each, with both modes running side by side: a story of about 2,000 words, a JavaScript Tetris engine of about 290 lines, and the numbers from one to four hundred written out in words. We took off the usual wait before a reply starts, estimated from a one-word reply in the same round.

What it wrote	Standard	Ultrafast
A story	30.9	139.6 (4.5x)
Code	32.5	282.9 (8.7x)
Numbers in words	33.1	490.7 (14.8x)

Those are tokens a second, the middle result of three rounds, from GPT-6 Astra at low effort in Codex on 30 September 2026. Ultrafast’s gain over Standard is in brackets.

Standard wrote everything at about the same pace, 28 to 34 tokens a second across all nine replies. In our tests, Ultrafast gained more the more predictable the text was. Its counting replies took only 5 to 9 seconds each, so the wait we took off counted for more, and they ranged from 305 to 718 tokens a second.

Even Standard writes faster than you can read: about 25 words a second on the story. Adults reading silently manage about 4 (238 words a minute for non-fiction and 260 for fiction, in Brysbaert’s 2019 review of 190 studies).

Why a whole job speeds up less

A job in Codex is more than writing. The model thinks before it writes, runs commands and checks its work. In our first take, both runs also tested the game in a browser. Ultrafast wrote its main file 10.5 times faster than Standard (342.5 tokens a second against 32.7), but its whole first version was only 3.5 times faster.

First take	Standard	Ultrafast
Thinking and waiting	62.2 s	40.2 s
Writing	294.1 s	39.8 s
Running commands and checks	37.7 s	31.2 s
First version done	396.6 s	113.6 s

Thinking and waiting runs from the start of each model call to the end of its reasoning, so it includes the wait before a reply starts. A few seconds of each run fell between steps, so the rows don’t quite add up. The runs also did different amounts of work: Ultrafast wrote 15,428 output tokens in all, and Standard 10,448.

This is Amdahl’s law. Gene Amdahl made the point in 1967, in a conference paper about computers that split work across many processors: “the effort expended on achieving high parallel processing rates is wasted unless it is accompanied by achievements in sequential processing rates of very nearly the same magnitude.” His paper has no equation. The usual form, as Hill and Marty set it out in 2008, is:

New time = old time × ((1 − p) + p ÷ s)

Here p is the share of the time spent writing and s is how many times faster the writing gets. However big s gets, the job can’t get more than 1 ÷ (1 − p) times faster.

In our first take, Standard spent 74% of its time writing (294.1 of 396.6 seconds), and Ultrafast wrote its main file 10.5 times faster. Put those in and the whole job should come out about 3 times faster: 1 ÷ (0.26 + 0.74 ÷ 10.5) = 3.03. We measured 3.5 times, because Ultrafast was also quicker at the other parts: 40 seconds thinking and waiting against 62, and 31 seconds of commands and checks against 38. Its writing across all its calls took 39.8 seconds, more than the 28 the formula assumes, partly because it wrote more, but the time it saved elsewhere more than made up for that. With 74% of a job spent writing, even instant writing couldn’t make it more than 3.8 times faster.

Our second take had almost nothing to run, under 3 seconds of commands in each run, because neither run opened a browser. Ultrafast also spent far less time thinking and waiting there, 17 seconds against 124, and its first version came out 9 times sooner. So treat the calculator’s answer as a rough guide. It counts thinking and waiting as fixed, and in both our runs Ultrafast finished sooner than it predicts.

How it might be this fast

OpenAI hasn’t said. Its docs call Ultrafast the fastest service tier, and its API changelog says to use it “to reduce the time between generated output tokens”.

The first Ultrafast, for an older model called GPT-5.6 Sol, runs on Cerebras chips. OpenAI previewed it in the API on 13 August 2026, saying it was “Powered by Cerebras” and generated “up to 750 output tokens per second”, up to 14 times faster than Standard. The API guide still lists it as a preview. Cerebras makes chips the size of a whole silicon wafer, with 44 GB of memory on each, so a model’s weights can stay on the chip. Cerebras says that version uses the same model, weights and precision as Standard: “The only difference is the hardware the model is running on”. Nothing OpenAI or Cerebras has published says what runs Ultrafast for GPT-6 Astra.

Two things are true of AI chips in general, though OpenAI hasn’t tied either to Ultrafast. At each step of writing, a model reads its weights from memory, and that memory traffic, more than the sums, usually sets the speed. And serving fewer people at once on each chip makes each reply faster but each token dearer. As Pope and others put it in 2022, “As we decrease the batch size, we improve the latency but incur higher cost per token.” That would fit Ultrafast costing 6 to 8 times as much.

Faster chips or smaller batches alone should speed up every kind of text by about the same amount. Our results didn’t: Ultrafast gained 4.5 times on a story and 14.8 times on counting. That pattern fits speculative decoding, a technique published in 2023 by Leviathan, Kalman and Matias, and separately by Chen and others at DeepMind. A small, fast model drafts the next few tokens, and the big model checks them all at once, keeping the guesses up to the first one it disagrees with. When the guesses land, it skips ahead. The Leviathan paper reports “identical outputs”, so the answer stays the same and only the speed changes.

Predictable text means more of the guesses land. DeepMind’s team found it sped up code (the HumanEval tests) 2.46 times, against about 2 times for summaries (XSum), and put that down partly to code “containing a lot of common sub-sequences”.

That’s our best guess, not OpenAI’s word. Fast chips and speculative decoding can work together, and OpenAI could be using something else entirely.

When it’s worth it

Ultrafast earns its cost when you’re sat waiting for the answer: trying out ideas, going back and forth on a change, or fixing something that’s broken now. Each of our six upgrades took between 15 and 27 seconds, short enough to wait for.

For a long job you leave running, it uses your allowance 8 times as fast, and you weren’t waiting for it anyway. Standard makes more sense there, or Fast at 2.5 times the usage if you want some of the speed. The more of a job is the model writing, especially code, the more Ultrafast saves.

These are our timings, with one model on one day. Yours will vary.

Sources

• OpenAI Codex docs, checked 30 September 2026: Speed, Pricing, Configuration reference, Slash commands and Models
• OpenAI API docs: Ultrafast mode, Changelog, Pricing and Key concepts (tokens)
• OpenAI, 3 September 2026: GPT-6 Astra launch post; 13 August 2026: Ultrafast preview for GPT-5.6 Sol
• Codex’s model list for GPT-6 Astra’s Fast mode: models.json in the openai/codex repository
• Cerebras, 13 August 2026: Accelerating GPT-5.6 Sol Ultrafast with OpenAI; 27 August 2026: How Cerebras serves GPT-5.6 Sol at up to 750 tokens per second
• Gene Amdahl, 1967: Validity of the single processor approach to achieving large scale computing capabilities, AFIPS Spring Joint Computer Conference
• Hill and Marty, July 2008: Amdahl’s Law in the Multicore Era, IEEE Computer
• Leviathan, Kalman and Matias, 2023: Fast Inference from Transformers via Speculative Decoding
• Chen and others, DeepMind, 2023: Accelerating Large Language Model Decoding with Speculative Sampling
• Pope and others, 2022: Efficiently Scaling Transformer Inference
• Brysbaert, 2019: How many words do we read per minute?
• Kettles: Which?, How we test kettles; Breville, IQ Kettle BKE820 (1,500 W, 110 to 120 V); Durand and others, 2022, Electric Kettles: An Assessment of Energy-Saving Potentials for Policy Making in the European Union (the 85% efficiency)
• Our timings: Codex CLI 0.159.2 with GPT-6 Astra on 30 September 2026, the race at medium effort and the single replies at low effort

This guide is based on what OpenAI and Cerebras had published by 30 September 2026, and our own timings that day.